Background
Twenty years of software engineering. My current work sits at the seam between machine learning models and the systems that host them: serving models on on-premise hardware, tuning latency and memory, and wiring them into existing application stacks. Before that I shipped consumer applications across Apple platforms and the web, with audio and graphics as recurring threads.
Polyglot: Portuguese, Romanian, English, Spanish.
Essays
Valentin Radu · 2026 ·
Valentin Radu · 2026 ·
Bookshelf
Les Misérables (Victor Hugo)
A book worth revisiting every decade. It reads differently each time.
The Feynman Lectures on Physics (Richard Feynman)
A fascinating account of how the world actually works, as much as we can figure it out anyway.
Friedrich Nietzsche
He looked at the world and wrote what he truly saw, with comprehension and without filter. A pillar of intellectual honesty.
Core Competencies
Applied machine learning
Self-hosted serving, GPU/CPU inference tuning, vector search and retrieval, codebook design
Low-level systems
Unix internals, concurrency primitives, hot-path optimisation
Application development
iOS and macOS native, web frontends, audio DSP, graphics and shaders
Experience
Founder
Kinoto · Porto · 2023 – Present
Technical writer
Swiftcraft.io · 2019 – 2026
Software engineer
Berlin, London, Hamburg · 2015 – 2023
Application developer
Cluj, Romania · 2010 – 2015
Education
Computer Science · Philosophy
Awards
Google Android Developer Challenge
Trip Journal · Grand Prize · 2009
Dev journal
npm install and pray. Axios and TanStack, 2026.
May 2026Two npm supply chain attacks in early 2026: the Axios compromise by Sapphire Sleet and the Mini Shai-Hulud worm by TeamPCP. What happened, how to check if you were hit, how to harden, and an honest look at how Pent would have mitigated these.
Vector search
Updated May 2026Nearest neighbor search at scale, worked from the geometric problem itself before the algorithms. Why exact methods break down in high dimensions, what approximate algorithms trade for speed, and how production indexes are actually built.
Show all 4 partsHide parts
- 1.Vector search - The problem (above)
- 2.Vector search - Small worlds
- 3.Vector search - On disk
- 4.Vector search - Quantization
AI and technical writing
May 2026Almost every public article online is written with AI assistance now. So are the articles on this site. A note on how I write now, and why.
I built a compiler frontend with Claude's help. A retrospective.
Mar 2026A writeup of building formalang, a compiler frontend, with Claude in the loop. The actual workflow, what it cost in time and reading, where the agent helped most, and a final scorecard.
Self-hosting Devstral Small 2 on a 48 GB GPU
Dec 2025Deploy Mistral's Devstral Small 2 (24B coding model) with vLLM on a 48 GB workstation GPU or rented cloud instance, then connect it to OpenAI-compatible agentic clients (Cline, Continue, OpenCode).
Rust for Swift devs
Updated Nov 2025A walk from Swift's mental model to Rust's, across the type system, error handling, memory and ownership, lifetimes, concurrency, and tooling. Aimed at developers comfortable in Swift who want a precise dictionary into Rust.
Show all 6 partsHide parts
- 1.Rust for Swift devs - The type system (above)
- 2.Rust for Swift devs - Error handling
- 3.Rust for Swift devs - Memory and ownership
- 4.Rust for Swift devs - Lifetimes
- 5.Rust for Swift devs - Concurrency
- 6.Rust for Swift devs - Tooling and interop
The exponential thread
Updated Oct 2023A nine-part climb starting from a 3rd-grade equation and ending at the exponential map of a Lie group. The thesis is that e, ln, sin, cos, complex exponentials, matrices, and rotations are projections of one underlying object, and the climb makes that statement concrete one rung at a time.
Show all 2 partsHide parts
- 1.The exponential thread - The puzzle (above)
- 2.The exponential thread - Why e is natural
AWS Lambda + Rust
Mar 2019Deploying Rust on AWS Lambda in 2026: how the toolchain has changed since 2019, and what a minimal, idiomatic setup looks like today.